AgentRelay

AgentRelay helps you switch coding agents without losing momentum.

It reads your current AI session, captures real project state (including git context), and generates a high-quality continuation prompt so the next agent can continue from where you actually stopped.

Problem

When you switch agents mid-task, context usually breaks.

• Important decisions are scattered in chat history
• Errors and failed attempts get repeated
• The next agent starts from an incomplete picture

Solution

AgentRelay builds a clean handoff prompt from:

• Session messages
• Tool calls and outputs
• Open errors and failed approaches
• Current git state and changed files

Result: faster context transfer, less repetition, better continuity.

Supported Sources

• Claude Code
• Codex
• OpenCode

Supported Targets

• generic
• claude
• codex
• cursor
• chatgpt

Features

• Smart latest-session detection for the current project
• Focused context extraction with noise reduction
• Error-aware prompt generation
• Git-aware summaries (status, diffs, untracked files)
• Optional prompt refinement via OpenRouter
• Clipboard copy on macOS, Linux, and Windows
• Built-in diagnostics with doctor command
• Session browsing with sessions command

Install

Run once with npx:

npx @rajdevxd/agentrelay

Install globally:

npm i -g agentrelay

Quick Start

cd your-project
agentrelay

You will:

1. Pick source agent (if not provided)
2. Parse latest session for current repo
3. Capture git/project context
4. Generate continuation prompt
5. Copy output to clipboard

Usage

# Basic
agentrelay

# Choose source
agentrelay --source claude
agentrelay --source codex
agentrelay --source opencode

# Choose target style
agentrelay --source claude --target codex

# Use exact session
agentrelay --source codex --session <id-or-path>

# Save output to file
agentrelay --output ./handoff.md

# Refine with OpenRouter
agentrelay --refine

# Pick preserved context start interactively
agentrelay --pick-user

# Start preserved context from nth user prompt
agentrelay --from-user 2

# Diagnostics
agentrelay doctor --source claude

# List sessions
agentrelay sessions --source codex --limit 5

Commands

• continue (default): generate handoff prompt
• doctor: validate environment, source availability, git, clipboard, and API key state
• sessions: list recent sessions for current project

Flags

Flag	Description
--source	Source agent: claude, codex, opencode
--target	Target format: generic, claude, codex, cursor, chatgpt
--session <id|path>	Use specific session
--pick-user	Pick starting user message interactively
--from-user	Start preserved context from nth user message
--output, -o	Write final prompt to file
--refine	Refine prompt with provider
--provider	Refinement provider (default: openrouter)
--model	Override model
--api-key	Override provider API key
--limit, -n	Session rows limit for sessions command
--help, -h	Show help
--version, -v	Show version

LLM Refinement

Refinement is optional.

• Provider: openrouter
• API key resolution order: CLI flag, env variable, local config

Environment variable:

OPENROUTER_API_KEY=your_key_here

Local config file path:

~/.agentrelay.json

Session Storage Paths

Source	Path	Format
Claude Code	~/.claude/projects//*.jsonl	JSONL
Codex	~/.codex/sessions////*.jsonl	JSONL
OpenCode	~/.local/share/opencode/opencode.db	SQLite

Requirements

• Node.js 18+
• One supported source agent with existing session data
• sqlite3 CLI only for OpenCode source
• Git repository recommended for best code-state capture

Troubleshooting

• No sessions found:
  • Run source agent once in the current project
  • Run agentrelay sessions --source
  • Run agentrelay doctor --source
• OpenCode parsing issues:
  • Ensure sqlite3 CLI is available
• Refinement failed:
  • Check OPENROUTER_API_KEY
  • Retry without --refine to use raw mode

Maintainer

• GitHub: https://github.com/codewithevilxd
• Portfolio: https://nishantdev.space
• Email: codewithevilxd@gmail.com

License

MIT